方法论 / 技术路线

国自然研究框架图（四研究层级）

面上项目申请书风格的研究框架图，多研究内容相互支撑，调节变量明示。

完整 Prompt

End-to-end training pipeline for a semantic segmentation model on satellite imagery. Five stages connected left-to-right by labeled black arrows on a white background.

Stage 1 — Input
Surface reflectance Landsat tiles (atmospherically corrected). Show a 2x2 grid of tile thumbnails inside the stage.

Stage 2 — Pre-processing
- Remove zero-value pixels.
- Slice large scenes into 512x512 tiles.
- Normalize per channel.

Stage 3 — Splits
Train (80%) / Val (10%) / Test (10%). Show as a horizontal stacked bar with each split labeled.

Stage 4 — Model
U-Net backbone with encoder-decoder skip connections. Show the U-shape inside this stage.

Stage 5 — Output
Per-pixel class probabilities -> argmax -> segmentation map. Show a small example mask thumbnail.

Style: clean rounded rectangle boxes, minimal palette (navy, teal, gray), sans-serif labels, no decorative icons. IEEE / Remote Sensing journal style.

立即试用此 Prompt

适用场景

For methodology sections of remote sensing, medical imaging, or any segmentation paper.

变体

Top-down, single-column variant

Same pipeline content, but laid out top-to-bottom in a single column, suitable for a vertical column figure in a 2-column paper. Each stage is a row; arrows are short vertical chevrons.

With augmentation branch

Add a sixth side branch between Stage 2 and Stage 3 labeled "Augmentation" containing rotation, flip, color jitter, and CutMix. Show the branch feeding back into the train split only.

使用建议

Number every stage. Numbered stages reduce label drift.
List 2–4 bullets inside each stage — more than that gets truncated visually.
Specify the exact thumbnail you want inside each stage (e.g. "show a 2x2 grid of tile thumbnails").

常见问题

Can I show two pipelines side-by-side for comparison?

Yes — say "Two parallel rows, top: baseline pipeline, bottom: proposed pipeline. Aligned columns highlight the differences."

How do I add a feedback loop from evaluation back to training?

Add "Dashed feedback arrow from Stage 5 (Output) back to Stage 4 (Model) labeled \"loss / gradient\"."